<html><body><div style="font-family: arial, helvetica, sans-serif; font-size: 12pt; color: #000000"><div><div>Research engineer or post-doc position in Natural Language Processing:</div><div>Introduction of semantic information in a speech recognition system</div><div><br></div><div><br></div><div> Supervisors: Irina Illina, MdC, Dominique Fohr, CR CNRS</div><div><br></div><div>Team: Multispeech, LORIA-INRIA (https://team.inria.fr/multispeech/)</div><div><br></div><div>Contact: illina@loria.fr, dominique.fohr@loria.fr</div><div><br></div><div>Duration: 12-15 months</div><div><br></div><div>Deadline to apply : March 15th, 2019</div><div><br></div><div>Required skills: Strong background in mathematics, machine learning (DNN),</div><div>statistics, natural language processing and computer program skills (Perl,</div><div>Python).</div><div><br></div><div>Following profiles are welcome, either:</div><div><br></div><div>· Strong background in signal processing</div><div>or</div><div>· Strong experience with natural language processing</div><div><br></div><div>Excellent English writing and speaking skills are required in any case.</div><div><br></div><div>Candidates should email a detailed CV with diploma</div><div><br></div><div>LORIA is the French acronym for the “Lorraine Research Laboratory in Computer</div><div>Science and its Applications” and is a research unit (UMR 7503), common to</div><div>CNRS, the University of Lorraine and INRIA. This unit was officially created in</div><div>1997. Loria’s missions mainly deal with fundamental and applied research in</div><div>computer sciences.</div><div><br></div><div>MULTISPEECH is a joint research team between the Université of Lorraine, Inria,</div><div>and CNRS. Its research focuses on speech processing, with particular emphasis</div><div>to multisource (source separation, robust speech recognition), multilingual</div><div>(computer assisted language learning), and multimodal aspects (audiovisual</div><div>synthesis).</div><div><br></div><div> Context and objectives</div><div><br></div><div>Under noisy conditions, audio acquisition is one of the toughest challenges to</div><div>have a successful automatic speech recognition (ASR). Much of the success</div><div>relies on the ability to attenuate ambient noise in the signal and to take it</div><div>into account in the acoustic model used by the ASR. Our DNN (Deep Neural</div><div>Network) denoising system and our approach to exploiting uncertainties have</div><div>shown their combined effectiveness against noisy speech.</div><div><br></div><div>The ASR stage will be supplemented by a semantic analysis. Predictive</div><div>representations using continuous vectors have been shown to capture the</div><div>semantic characteristics of words and their context, and to overcome</div><div>representations based on counting words. Semantic analysis will be performed by</div><div>combining predictive representations using continuous vectors and uncertainty</div><div>on denoising. This combination will be done by the rescoring component. All our</div><div>models will be based on the powerful technologies of DNN.</div><div><br></div><div>The performances of the various modules will be evaluated on artificially noisy</div><div>speech signals and on real noisy data. At the end, a demonstrator, integrating</div><div>all the modules, will be set up.</div><div><br></div><div>Main activities</div><div><br></div><div> • study and implementation of a noisy speech enhancement module and a</div><div> propagation of uncertainty module;</div><div> • design a semantic analysis module;</div><div> • design a module taking into account the semantic and uncertainty information.</div><div><br></div><div>References</div><div><br></div><div>[Nathwani et al., 2018] Nathwani, K., Vincent, E., and Illina, I. DNN</div><div>uncertainty propagation using GMM-derived uncertainty features for noise robust</div><div>ASR, IEEE Signal Processing Letters, 2018.</div><div><br></div><div>[Nathwani et al., 2017] Nathwani, K., Vincent, E., and Illina, I. Consistent DNN</div><div>uncertainty training and decoding for robust ASR, in Proc. IEEE Automatic</div><div>Speech Recognition and Understanding Workshop, 2017.</div><div><br></div><div>[Nugraha et al., 2016] Nugraha, A., Liutkus, A., Vincent E. Multichannel audio</div><div>source separation with deep neural networks. IEEE/ACM Transactions on Audio,</div><div>Speech, and Language Processing, 2016.</div><div><br></div><div>[Sheikh, 2016] Sheikh, I. Exploitation du contexte sémantique pour améliorer la</div><div>reconnaissance des noms propres dans les documents audio diachroniques”, These</div><div>de doctorat en Informatique, Université de Lorraine, 2016.</div><div><br></div><div>[Sheikh et al., 2016] Sheikh, I. Illina, I. Fohr, D. Linares, G. Learning word</div><div>importance with the neural bag-of-words model, in Proc. ACL Representation</div><div>Learning for NLP (Repl4NLP) Workshop, Aug 2016.</div><div><br></div><div>[Mikolov et al., 2013a] Mikolov, T. Chen, K., Corrado, G., and Dean, J.</div><div>Efficient estimation of word representations in vector space, CoRR, vol.</div><div>abs/1301.3781, 2013.</div><div><br></div><div><br></div><div>Associate Professor</div><div>Lorraine University</div><div>LORIA-INRIA</div><div>office C147</div><div>Building C</div><div>615 rue du Jardin Botanique</div><div>54600 Villers-les-Nancy Cedex</div><div>Tel:+ 33 3 54 95 84 90</div></div><div></div></div></body></html>